Slurm Hpc

slurm-hpc — generic Slurm workload-manager know-how

Portable Slurm reference for authoring batch jobs and reasoning about resource isolation on any cluster. Not tied to a specific site — for a specific repo's recipes and layout, prefer that project's own skill.

When to use

• Writing or fixing an sbatch script, or an srun command line.
• Choosing resource requests (CPUs, memory, GPUs, time, partition).
• Reading job/queue/node state (squeue, sacct, sinfo, scontrol).
• Reasoning about what isolates jobs (cgroups) and what fences GPU VRAM.

When NOT to use

• Operating a specific project's cluster with its own recipes → use that repo's skill.
• Designing cluster provisioning/config from scratch → that's admin work, not this.

Batch script skeleton

#!/bin/bash
#SBATCH --job-name=train
#SBATCH --partition=<partition>       # see `sinfo`
#SBATCH --gres=gpu:1                  # GPUs (see GPU section)
#SBATCH --cpus-per-task=8
#SBATCH --mem=32G                     # per-node; enforced as a hard cgroup cap
#SBATCH --time=02:00:00               # job killed at this wall-clock limit
#SBATCH --output=logs/%x_%j.out       # %x=name %j=jobid
srun python train.py

Submit sbatch script.sh; inspect squeue -j <id>; after it ends, sacct -j <id> --format=JobID,State,ExitCode,Elapsed,MaxRSS,ReqTRES,AllocTRES. Cancel scancel <id>. --start on squeue estimates a pending job's start.

Resource requests

• --mem=<N>G / --mem-per-cpu=<N>G — real memory; with cgroups this is a hard cap (see below). --cpus-per-task + --ntasks set core count.
• --gres=gpu:N or --gres=gpu:<type>:N (e.g. gpu:a100:2).
• --time is mandatory discipline on most clusters; jobs are killed at it.
• --exclusive gives the whole node. Partitions (sinfo) gate limits/priority.

Chaining and waiting

Do not babysit a job by re-running squeue on a timer. Either hand the sequencing to the scheduler, or block on it once.

Chain B behind A — Phase B is queued at submit time and runs without anything staying connected:

JID=$(sbatch --parsable phase_a.sbatch); JID=${JID%%;*}   # strip ";cluster"
sbatch --dependency=afterok:"$JID" \
       --kill-on-invalid-dep=yes \
       --mail-type=INVALID_DEPEND,END,FAIL \
       phase_b.sbatch

Separators: , means all must be satisfied, ? means any.

Block until done (one command, exit code mirrors the job's):

sbatch --wait phase_a.sbatch

Checkpoint before the wall-clock kill — --signal=B:USR1@300 sends SIGUSR1 to the batch shell 300 s before --time expires; trap it and save. The B: prefix is what targets the batch shell rather than the job steps.

For the agent-side question — how should I, the agent, wait for this? — see the long-running-jobs skill, which ranks scheduler chaining, one blocking backgrounded wait, filtered event streaming, and scheduled check-ins.

Isolation: what actually fences a misbehaving job

CPU/RAM (cgroups, if the site enables task/cgroup + ConstrainCores/RAMSpace): a job is pinned to its cores (cpuset) and hard-capped at --mem. Exceeding --mem triggers the kernel cgroup OOM killer inside that job's cgroup only — the job dies alone; neighbours and the host are untouched. A thread-bomb only saturates its own cores. This is real isolation and is the norm.

GPU VRAM — the gap. Allocating a GPU does NOT cap its memory:

Read references/gpu-isolation.md when the task is specifically about capping or partitioning GPU VRAM, or choosing between shard/mps/MIG.

Reference files

• references/gpu-isolation.md — read when capping/partitioning GPU VRAM or choosing shard vs mps vs MIG; has the config snippets and enforcement details.

Gotchas

• --gres=shard:N and plain --gres=mps:N do NOT fence VRAM. Sharing a GPU ≠ isolating it. Only MPS with a pinned memory limit set, or MIG, enforces a per-job VRAM ceiling. A job with a wrong batch size otherwise OOMs the whole card and takes down its neighbours.
• MIG needs a datacenter GPU (A100/A30/H100…); GeForce/RTX cards can't do it.
• --mem is a hard cap under cgroups, not a hint. A job that needs more gets OOM-killed; bump --mem, don't expect bursting into idle RAM.
• sacct needs accounting (slurmdbd) configured. Without it use scontrol show job <id> (live only — gone after the job ages out).
• afterany is the default dependency type. A bare -d 12345 runs the child after the parent terminates either way — including after it crashed. Always spell out afterok:.
• A failed parent leaves the child PENDING forever. Slurm's default is "the job stays pending with reason DependencyNeverSatisfied", which in squeue looks identical to waiting for resources. Worse, "the dependent job will never be run, even if the preceding job is requeued" — fixing and requeueing the parent does not release it; you must resubmit the child. Pass --kill-on-invalid-dep=yes and --mail-type=INVALID_DEPEND.
• --parsable prints jobid;cluster, not a bare id, when a cluster name is configured — "The values are separated by a semicolon." Strip it with ${JID%%;*}, or the dependency string is malformed and the ; truncates your shell command.
• sbatch --wait collapses every signal death to exit 1. "If the job terminated due to a signal rather than a normal exit, the exit code will be set to 1." An OOM kill, a TIMEOUT, a scancel, and a plain exit 1 are indistinguishable from the exit code. Read sacct -j <id> --format=State,ExitCode for the real state — the :signal suffix and states like OUT_OF_MEMORY / TIMEOUT / NODE_FAIL are what tell you whether a retry could work. For arrays the code is "the highest value for any task".
• A failing site Prolog/Epilog drains the node, showing jobs as PENDING with a node Reason=. Check scontrol show node.
• srun --oversubscribe is ignored under consumable resources — you can't opportunistically grab more than you requested.